CoFAE-LLM framework: LLM-based diagnosis framework for multi-view evidential reasoning on rolling bearing vibration signals

Abstract To address the limitations of single-representation learning and the lack of credibility modeling in multi-view fusion for rolling bearing fault diagnosis, we propose CoFAE-LLM, an LLM-based diagnosis framework for multi-view evidential reasoning on vibration signals. First, the multi-view evidential layer constructs time-domain (time-view), time–frequency (TF-view), and statistical-view features from raw vibration signals, and parameterizes them as Dirichlet evidence in a shared label space, yielding comparable class-support strengths and calibrated predictive uncertainty across views. Second, the relation-aware multi-view evidential fusion module models inter-view relationships in the evidential space and estimates sample-wise reliability, thereby discounting low-credibility evidence and performing robust aggregation to obtain fused evidence and its relational context. Subsequently, an evidence-driven soft-prompting mechanism maps the fused evidence and relational context into continuous prefix embeddings, which conditions the LLM to perform diagnostic reasoning under evidential constraints. Experiments on three public datasets show that CoFAE-LLM accuracies of 99.27%, 98.64%, and 98.11% on CWRU, SEU, and JNU, respectively. Relative to the strongest baseline, WDCNN, CoFAE-LLM improves Accuracy/F1 by 1.02/0.80% points on CWRU and 1.22/1.22% points on JNU, while remaining competitive on SEU. Ablation results further confirm the effectiveness of the key modules.